Continuous Cardiac Arrest Prediction in ICU using PPG Foundation Model

TL;DR

This work tackles in-hospital cardiac arrest prediction in ICU using only unimodal finger PPG signals. It introduces FEAN, a two-stage architecture that leverages pre-trained PPG foundation models (PPG-GPT up to 1B) and a sequence aggregator to produce prognostic embeddings. Evaluated on UCSF ICU data with 200 positive and 1000 negative cases, it uses 1-hour (1H) and full-history (FH) variants over a 24-hour prediction horizon, achieving average AUROC around 0.79 and peak 0.82 one hour before CA; PaCMAP visualizations reveal latent trajectories reflecting health deterioration. The results demonstrate the viability of unimodal PPG foundation models for critical-event prediction and motivate cross-institution validation and time-to-event modeling in future research. Overall, the study suggests that large-scale PPG foundations can approach or match multi-modal methods while enabling scalable, non-invasive monitoring in ICUs.

Abstract

Non-invasive patient monitoring for tracking and predicting adverse acute health events is an emerging area of research. We pursue in-hospital cardiac arrest (IHCA) prediction using only single-channel finger photoplethysmography (PPG) signals. Our proposed two-stage model Feature Extractor-Aggregator Network (FEAN) leverages powerful representations from pre-trained PPG foundation models (PPG-GPT of size up to 1 Billion) stacked with sequential classification models. We propose two FEAN variants ("1H", "FH") which use the latest one-hour and (max) 24-hour history to make decisions respectively. Our study is the first to present IHCA prediction results in ICU patients using only unimodal (continuous PPG signal) waveform deep representations. With our best model, we obtain an average of 0.79 AUROC over 24~h prediction window before CA event onset with our model peaking performance at 0.82 one hour before CA. We also provide a comprehensive analysis of our model through architectural tuning and PaCMAP visualization of patient health trajectory in latent space.

Figures (5)

• Figure 1: Illustation of Feature Extractor-Aggregator Network (FEAN). A pre-trained foundation model extracts features and a sequence model aggregates the embedding sequence into a single vector which finally predicts the target.
• Figure 2: Illustration of the overall pipeline (for a case patient). 1-hour model ("1H") and Full-history ("FH") model are trained on 1 h and all history (limited to 24 h max) before a chosen anchor timepoint in $[T-25, T-1]$ resp. During the evaluation, predictions are made every one hour (alarm frequency) i.e. at T-24, ..., T-1 where t=T is the time of event onset time.
• Figure 3: Illustration of moving average of PaCMAP mapping of PPG-GPT embeddings of 1 hour PPG data before cardiac arrest event. Each time point corresponds to a 30-second segment. Note that the trajectory follows a defined path, potentially capturing the patient's health trajectory in the latent space.
• Figure 4: Hourly AUROC for STFT baseline and 1H 345M (both feature extractor tuned and frozen). For the feature extractor-tuned model, overall performance is best (printed in legend) and improves when getting closer to the event onset time.
• Figure 5: Receiver Operating Curve (ROC) values averaged over 24-hour prediction window. Here, prevalence is 11%. The optimal threshold will depend on metrics to minimize, for instance, false alarms to reduce alarm fatigue hu2012predictive.